Connor Zwick: Making Language Immersion Possible Through AI

24:43And Berkeley was, I think one of the first like real, I mean, there was like some super early like machine learning stuff, of course, but not like the interesting cutting edge stuff. And we found a course at Berkeley, it was like a graduate level course. It was about reinforcement learning, which was like the newest, like kind of most interesting. Can you just explain it for listeners, reinforcement learning? Yeah. So, you know, if people are following Gen AI, they, you know, may be, they may have heard of this concept of like rlhf reinforcement learning human feedback um and basically the idea is that this was like um like when people were you know uh releasing models to like you know beat games alpha go um uh or like starcraft or any of this stuff this is what reinforcement learning is it's essentially like um strategy-based intelligence so it's like uh making the correct move over time and kind of keeping a strategy in mind to execute over time and doing and like basically self-learning that.

25:41So like, just like trying a bunch of times until you get like the rewards. And so that's why it's called reinforcement learning. Anytime you do a good thing, you say good robot and you give it a reward. Anytime it does a bad thing, you don't give it the reinforcement. And so very much like inspired by how the brain works and how humans learn. So this was showing a lot of really interesting promise. It was very intuitive that it would become a big, like maybe the path to AGI. And so a lot of people were focused on this. it ended up kind of mattering with like large language models because RLHF was a big unlock.

26:10And so OpenAI, even though they're, well, anyway, it's a tangent, but OpenAI like was super interested in RL in the beginning, everyone was, and then pivoted to LLMs, but like the RL stuff actually ended up being very useful still. But anyway, yeah, we found this course in 2015, back to the story. And we were like, what if we just show up? So it like had just started, I think it was like September of that year. And I remember going to Berkeley and we like take the BART there, me and Andrew. We have our backpacks on, we have like our notebooks and we're literally walking. I think it's like, it was like the first week of school because it was like the first lecture.

26:47And we're getting stopped by like an undergrad. And he was like, hey, my name's, you know, whatever. And he introduces himself and he's trying to make friends because he thinks we're also just like these undergrads going to school. And we felt like such imposters because we literally were. and we end up like getting into the building and finding the classroom and we were expecting a giant lecture hall and it's literally like 12 people. And it's like a graduate level course about this like pretty specific topic. And so there's no one in the course. It's like no one had ever even really heard of reinforcement learning.

27:19They didn't even know what it was. And so we just like walk in and sit at the back of the classroom. Everyone's staring at us. It's very clear. Everyone kind of knows each other from other courses and they're like, who are these people? But everyone's a little too awkward to actually like confront us. And including the, I think the lecturer definitely knew for sure. And yeah, it was, so we just literally would go every single time. So the lecturer ended up being John Schulman, who ended up being one of the co-founders of OpenAI. He like, he was the one teaching it, even though Peter Abiel was like the actual professor, But like, I'm sure Abia like had less, less knowledge of reinforcement learning because it was such a new topic than John Schulman did, who was like, I think grad, a grad student at that point, maybe, maybe.

28:06But anyway, like, yeah, it was like, literally, I'm sure the people that were in that room, all if I knew who they were, like, they probably have all gone on to like join research labs and such. Did he ever confront you? Did he ever, did he ever stop? Hey guys, I know you're not in this course. What are you doing here? Yeah, never. Did you ever have to submit any work? We so everything was on a public, everything was on a public website. And so it was amazing. He would just give us the website and everything was there. He also was, I think there was a period where it was literally three or four weeks in a row, where he was literally making the courses we went, and he like, was too busy.

28:42And so he didn't create homework. And so there literally was no homework for like three or four weeks. And then he finally reveals this thing. And he's like, Okay, this is going to be your new environment where you're going to like build your models, and then it'll test on this. And I'm pretty sure this was like the first version of the OpenAI playground or gym, which is like their big reinforcement learning thing that he ended up doing at OpenAI. And he literally like spent four weeks on it to do our like to basically give us like a homework environment to build models in like, you know, casually producing this for his students.

29:12So it was pretty cool. Did you ever tell him the story? I don't. I actually don't. I'm actually pretty good friends with his his wife now. I don't think I have. It'll be hilarious. At some point I will. That'll be a great, yeah, that'll be an amazing conversation. I'd be very curious if you remember those two randos in the back of the classroom. That's incredible. Okay, so you learn everything you need to know about AI by sitting in a Berkeley class that you had no business being in. Then what? Yeah, everything was published on Archive. It was such an open time in terms of all these papers being published.

29:45Basically, what we did was our entire strategy was we want to just implement as many papers as possible, build and learn by doing. And we eventually started building speech models. And so like one thing that I think was very, very, very important for Speak was we were building a speech recognition model. And while we were building it, we ended up deciding why don't we try to like actually understand more than just the words that people are saying. And also, people are like speech recognition models today are super bad for people speaking with accents or in loud environments. Like they basically were not like they were not accurate whatsoever, like basically not useful.

30:30And we were like, what if we try to intentionally find accented data on YouTube? We'll find like the BBC, for example, all that's in British English. We'll try to find stuff in like Indian dialect, English, et cetera. And maybe we can make the model robust by including all this data. Um, and we ended up building a model that could detect what accent people were speaking with and was incredibly robust to accents as well. And it was like basically state of the art and it was like a weekend project we did. Um, and I think for that, that was like a very kind of, you know, holy shit moment for us because the results were like mind-blowingly good.

31:09Um, and it really laid, laid bare to us. It was like probably the first thing that we did that was somewhat novel and it laid bear to us that these things are really, really powerful. And we're clearly like people are only, we're only mining the surface of what these things can do. And in our minds, I think it became very clear that it was, if you could extrapolate over time, it was merely a function of the amount of data that you have and like the size of the model and the compute. And invariably there'd be like, if like ways to make the models be even like more efficient for training and everything.

31:42but as you do that the accuracy of the model goes up and up and up and up like this and we have like the entire internet to train on and so it is just a matter of time until these models get to the level where at some point they're like better than a human right um and at the time at that at that point that's like an inflection because for all these models they're doing tasks that right now you know fundamentally the only thing like machines can't do and so as you approach human you start to become really useful as like a helper tool, right? Like, it's like an assistant tool, and the way that like co-pilot helps people, you know, write code, but it's not going to really write a program all by yourself.

32:21But once you surpass, and that's super valuable by itself, but once you surpass humans, that's where it's really a true unlock. And we became convinced that like, this is just a matter of time. And the kind of like, mantra that we've had from, from day one is basically that like, in the next five to 10 years, we'll be able to replace humans. for language learning. But that's kind of how we think about most things. It's, it was basically just like, it's only a matter of time. There's an inevitability here. These models clearly have a lot more juice to squeeze. And, um, I think that's like where everything started from was this like technological vision and really just like a complete conviction.

32:59Like it was, it wasn't just a vision. We knew this was going to be the case. Um, it was very, very clear. Um, like that, that models would just continue to get better. I mean, um, maybe not like AGI on a certain timescale, But like, but like the models would, at least in narrow cases, surpass humans. And that was really like the underpinning for speak. Once we knew that, once we knew that we thought at least speech models could get to the point where they were eventually superhuman, we could see that like all the pieces would unlock to do a lot of things. And the thing that we kind of immediately got to was language learning.

33:32We envisioned like we want to build, we were super obsessed with the idea of like, okay, if in five years we have a model that's sufficiently good, what is the thing that we want to be building? What feels like the most magical? And we wanted to make sure that it was something that humans could form a relationship with, something that felt really magical and that could really help them become better versions of themselves. For a variety of reasons, we got really focused on language learning. And that was kind of like the way we started Speak. We knew absolutely nothing about language learning, but we were convinced that over time, there'd be this thing that would be unlocked and we'd be able to help literally millions of people improve themselves in this way that really like changes people's lives.

34:12So I want to drill down into an insight you had before we dive into the speak story, which is you mentioned that back then, and by the way, was this 2015, 2016 when you're having this realization about the data? Yeah. So it feels like post chat GBT, you know, there's this, there's this big realization that if you feed these models enough data, they do exactly what you just said, right? And that seems to be a bit of an insight or a breakthrough moment that I feel like a lot of people have had over the past year. It sounds like you guys could tell this was going to happen in 2015. Am I understanding that correctly?

34:47And do you feel like there were a lot of people like you that knew that this would happen with enough data and enough time? It wasn't like a secret. No, no, no, no. This was not a secret. This is not just us. Like, like, it was just a smaller group of people, right? That knew, I think people that like less people were paying attention. But I mean, when we did YC, we so we did YC for speak again, second time. And that was winter 2017. By that point. So we started in 2015 through 2016 started speak at some point in late 2016, and then actually did YC. By that time, there was already a mini AI hype cycle that was happening.

35:24Like there were there there were a fair number of AI companies that existed in that batch. And there were VCs that were really interested in AI and it was very hyped. Like I remember a lot, the most, one of the most common questions we got was like, well, you're not actually trying to like teach people languages though, right? This is a Trojan horse to collect data so that you can like build models and then offer an API to other companies and do a B2B angle, right? And we were like, no, like we actually just like think that when this happens, we can unlock like a consumer use case that like billions of people want.

35:55And we think that'll be a really big company. And they kind of looked at us like we were crazy. But like there were a lot of people that were interested in like the data plays to unlock big AI models. And that was not like a super, I mean, I think we had, the unique thing about us is we had a lot of conviction to persevere, even though, because we really did understand it wasn't going to happen tomorrow. It was going to happen in five to 10 years. And we needed to like climb the steps to get there over time. So we'd be in a really good position to continue to build towards the long-term vision.

36:23And so I think a lot of these companies, like an interesting question is like of this earlier AI like hype cycle, like what big companies ended up emerging? And I think almost all the value, if you think about it, went to the incumbents like Google, like, you know, they were able to like the YouTube algorithm is so much better now because of deep learning. It's just like one example, ad targeting, huge unlock, like literally, you know, billions, tens of billions of dollars of value, probably. So there are lots of unlocks, but it all went to incumbents. And maybe there were a few like, you know, pick and shovel companies like scale that emerged, but there weren't many.

36:59But it seems like to your point about the YouTube algorithm and, you know, I can think of other companies that have had similar breakthroughs. It's probably more on, like you said, the sort of deep learning side, less on more of the generative side, which has really emerged over the over the past year or so. Is that fair to say? I mean, I guess what's really the difference is, I mean, I consider generative AIs. The creation. Just the buzzword. Yeah, but I'm thinking about speak, how it's talking back to you, right? Or I'm thinking about mid-journey or these things that are actually like creating content and experiences.

37:32I don't know. Maybe I'm mixing the two. I mean, I think it's a little bit nuanced. There's clearly, I think, something different now. But I think really what's happening is that all these models are like the transformer was created. The attention is all you need paper, I think, came out in like 2017. something like that. And so a lot of these models, like, you know, they've like all of this has existed. I don't really like think of that. Like, for example, Whisper is superhuman level speech recognition. That's transformer architecture. And it's just generating a transcript given an audio input, right?

38:09At any, any model is fundamentally generating something. It might just be like a classification or a recommendation, but it is like still generating some sort of output from an input. But and so I always like I'm a little bit amused by the like the fact that we use we're using this term generative. I've like I've started using it as well just because it's the term. It's like the parlance now at this point. But like I don't really think I think the big unlock was more just like that. These models got really good. Like GPT-3 came out in 2020 or 2021. It just wasn't as good because it was like DaVinci 1 or like Ada before that.

38:44And the thing is, these models are just continuing to get better. The curve is starting to become even sharper over time, right? And I think, like, the only thing that's happened is, like, basically, it's continued to get better to a certain point that unlocked value. And really, more than anything, ChatGPT proves that the UI mattered, right? Like, ChatGPT was just the DaVinci 2 model, which was available for, I think, nine months before ChatGPT came out. anyone could have built chat GPT basically. Um, if they had just like released that, no one did. And then they did and everyone paid attention.

39:22Um, and so I, yeah, I think really the function for me is like, it's the same thing. It's like everyone was working on LLMs have been around for, you know, at least three, four years at this point. And it's just a matter of like getting above a certain like utility level, I think. So you decide you want to do something in speech, but you don't really know what yet. Give us the story of speak from sort of the beginning to product market fit. And when did that happen? You know, we talked about product market fit earlier. You now have the perspective that you know when it's real and when it's not.

39:55When did you know it was real? And give us the story to get there. Yeah, well, it definitely took a few years of grinding. We knew nothing about language learning. I think even to back up a little bit further, we had this technology story in our heads. But I think one another thing that was super helpful, that's a very like a very specific memory for me is that Andrew and I spent an afternoon and we just posed the question in five years from now, what kind of company do we want to be running? At that point, we decided we think the best way to do this is not to go into research, but to start a company.

40:26Had you started the company yet? I don't think so. Yeah. This is probably pre starting the company. Yeah. Got it. OK. Still just an idea. OK, got it. Yeah. When we were doing research, we didn't even know if we wanted to start a company necessarily. We were intentional about like the best, like for our given motivations, we weren't necessarily sure that we wanted to do a company. But yeah, basically like we were like in five years, what would be like the most kind of incredible kind of company or position to be in? And we became really captivated by the idea of building something magical, something that really felt like novel and magical.

41:01And a big part of us for that was like something that you could actually interact with. And it would be like another human in the sense that it would feel like it understood you and you were going back and forth in some way. And we felt like it was clear that like chatbots in general were a ways off and it was less clear to us that that would be five years from now. But we felt like, and also when that happened, it would be something that like a giant tech platform company would probably do and not us. We didn't think that we would be the ones to be able to do that. Yeah. And also it was just like such a general problem space that it felt like they would be in a much better position to do it.

41:43And we always knew it would be like the hardest problem to do. Like the bar would be that much higher to be able to do that. However, like one thing we realized was like language learning is actually the perfect thing here because it's actually okay if the model isn't perfect. It's okay if it makes mistakes. And people are willing to talk with it anyway, because it's way better than the alternative, which is not having anyone to talk with, except for like a human, which is judges you and is inconvenient and is expensive. And we realized it's like one of the like, we thought about like what the categories could be of like what people want to interact with.

42:16We always we knew we wanted to do something more like straight consumer, we knew we didn't want to build for enterprise for this. And it's actually pretty limited. Like we pretty clearly identified early on, like language learning is one of the few things that adults do where they're willing to interact with a system like this and have it not be perfect. And it's one of the only areas of self-improvement and like especially learning that adults do in mass in general. And it was perfectly aligned for what we saw as like the speech recognition would get to human level faster than the language side would, if that makes sense.

42:53What do you mean not perfect? Like it's teaching you the wrong things or just the language that it speaks to? I mean, let's start with the speech recognition. Let's start with the speech recognition. Our first our first wedge here was the speech recognition. We knew we could do that well and better than better than what Google could provide or anyone else. Because the whole idea was we can get all this accented data from our users and then use it to train a model that's specifically good at like understanding them. And so let's just start there. But like even if we don't pick up every single word the user says, perfectly all the time, they'll still use the product.

43:26That was a hypothesis. Got it. Because it'll still be useful. But like if a chatbot like says kind of crazy things every once in a while or it doesn't really like understand your words all the time, like you're just not going to use that. The bar is too high. It's not like a finance application or something like where the stakes are really high. The stakes are high. We're just having a conversation. It's fine if it's not perfect. OK, exactly. Yeah. And so it's actually one of the few cases even today where you have, back to what I was saying earlier, like there's a difference between like an assistant tool where the human is in the loop, right?

43:57Copilot, or even for like legal, right? Like it's gonna be a very far time from now where we'll trust a legal review that's done by AI and there's no human in the loop, right? The stakes are way too high and mistakes are costly. And so there's actually very few use cases, even today where you can have the AI fully replace the human with no intervention from a human whatsoever. Because for most things, the stakes are just too high. And so chat is one of those. I think language learning is another. I don't think there are that many right now. When that happens for other industries, it'll be a really big deal for those industries, too.

44:34But I think that the hallucination rate and like a variety of other technical challenges are still too great that it hasn't happened yet. And I think that's like one of the most significant kind of like trends. Got it. Got it. That makes a lot of sense, actually. And it's really good context and framing for what can be a viable AI application in 2023 or 2024. OK, so, you know, you want to build a speech product because you feel like there's a tolerance for mistakes. Then what? Did you know what you wanted it to be at that point? We had, you know, very little of a vision around the specific way that we were going to teach people.

45:12like knew very little about language acquisition. And so, no, we started from scratch on all that. And that was a huge challenge for us. I think there were a few points where we narrowed it down. Like one thing we fundamentally knew from very early on was we wanted to focus on speaking first because it was the area that was most important to people and the least supported. There are already apps like Duolingo that helped you with vocabulary building, basic grammar, all that. That was covered. That's been covered for a long time. it's like less way less valuable. But speaking wasn't. And it was also the area where we could build the best models to actually like replace that that piece and actually get people speaking.

45:49The other thing we realized is that we needed some sort of like long term vision where we could build steps along the way and continue to improve our product market fit. And it was still valuable in the intervening years. We did not want to research project for five years until the models were good and then we're just going to go release something. We knew that like getting actual product market that was tough, it was going to require a lot of iterations. And we just wanted to like launch things and rapidly move. And like maybe the product market fit in the beginning wouldn't be as strong, but we knew that over time it would just, we would, you know, be with a rising tide.

46:19That's probably one of the things that you learned from flashcards that you need to be shipping product and getting feedback. Otherwise, like, I mean, so many first-time founders just build stuff for years before they ship it, but that feedback is so important. It's something I learned with a controller as well. And it was a limitation with hardware, Even though I knew I needed to be doing that, I couldn't do it because it was hardware. So, yeah, 100 percent. And then, yeah, I think like from there, we quickly honed in on the idea that like, OK, it's a vitamin for like people like us that are native English speakers that want to learn Spanish to go to Mexico City or something.

46:51But people that are trying to learn English. Wow. You don't realize that living in, you know, a place where most people are speaking English. But if you don't live in an English speaking country, it is like such a common need and it is an absolute painkiller. And man, there are way more of those people in the world than the other way around. So we quickly honed in on teaching English as well. That is a brilliant insight. Maybe it's such a subtle thing, but especially given that, you know, so much of tech, Silicon Valley, obviously based in the U.S., you naturally think of products for people based in the U.S., right?

47:25Like you said, are going to Mexico City. And clearly that's where products and companies like Duolingo have been very, very successful. Right. Maybe it seems obvious, but like, how did you have that insight? It's a simple one, but it's so powerful. I think like very early on, we realized like, even for me and Andrew, like we realized, oh, we're not that motivated to learn a language. Like, okay, do we know people that are? And so the motivation piece came up very early. And then I think we just naturally, as we were thinking about the market, we realized like, wow, but what about the people that are in the US that need to learn English?

47:57OK, well, what about like, oh, the fact that English is the lingua franca? It's the way that a Chinese person speaks to a Korean person as they speak in English. And then we start to really uncover it. It's a total blind spot for a lot of people in the Western world, especially the English speaking countries. and so yeah I mean I think that was a really big insight and you know also at the same time like a very hard pill to swallow because all of a sudden we were operating this meant operating in markets where we didn't speak the language we weren't familiar with the culture we weren't we didn't even like understand how to grow an app in these other markets and it was like a cerebral insight but it was very like difficult to operationalize it, if that makes sense.

48:38Yeah. And it was definitely like something that was required, like a lot of like commitment again to the idea that like long term, this is the right strategy. Yeah. And then from there, we realized, okay, even if we're doing that, now we're talking about teaching to a Chinese speaker, a Korean speaker, a Spanish speaker, a French speaker, you know, a Hindi speaker, right? Like all of these different language groups that were so vastly different. How do we start? And so we ultimately, we could go into it if it's interesting, but we ultimately decided we're going to even go more specific than this.

49:10We're going to just pick one country and one language, and we're going to get really, really zoomed in to that place. At this point, we've been struggling to find product market fit for a while, I should add. How long? How long is this journey to be like, hey, we need to start in other countries and other languages? I think probably it took us like six months to like even understand what like a early iteration could look like that made sense. And like testing a bunch of bad ideas first. And then we realized, you know, pretty shortly after that, OK, we need to teach English. And then it took us probably another three or six months to realize, OK, we need to go small.

49:46We need to go specific. We need to go pick a country. And then it took another year to find a profit. So we were grinding for years. Um, but yeah, we ended up picking, um, we ended up picking South Korea. Um, how do you do that? Like, I mean, yeah, I wouldn't even know where to start. Yeah. Yeah. It was, uh, well, uh, yeah, I think like it was a combination of many things. Um, this is probably one of the most common questions I get from like everyone from prospective employees, from, um, investors, from random people. Um, it was a combination of the fact that number one, um, we had a product out in many markets.

50:25And we saw that Korea is one of the places where people were using it in really interesting ways. And so there was natural, like this population of users, clearly there's something going on here. Let's take a closer look. I also ended up literally just traveling to a bunch of different countries and talking with our users for this early product. And Korea was one of those countries. And the thing that immediately stood out when I was in the country and I had done the research, right? We saw that like a crazy statistic that South Korea spent 1 % of their entire GDP on learning English. Um, and all of this like stuff around like the national obsession with it and the idea that no one, but it's still like one of the lowest, like English speaking ability, uh, countries.

51:05Um, but really, I think one of the things that cemented it for me was the fact that when I was driving down a street in Gundam with all the big skyscrapers, um, my, my friend who was kind of my tour guide was pointing out different skyscrapers and was like, that is a English learning Academy. That's an English learning Academy. Every single one of those windows you see in this giant skyscraper, those are all English classrooms. There are people here that are going after work or after school and they're studying for three hours late into the night every single day, English. And I was like, okay, this is crazy.

51:37This market is really interesting. And we realized like, okay, hyper-competitive market, lots of noise out there, lots of different English solutions and companies and products, super noisy. But if we can build something that's truly technologically differentiated, true product differentiation, right? That like that proverbial 10X, then we'll get a really good signal here. Because we also know people are really eager to actually get success here and they're willing to try anything. And so we'll get that market pull more here than anywhere else. We'll get the signal we care about for this long-term product market fit, which is the only thing that matters here more than anywhere else.

52:16And so that's basically how we did it. I can't imagine like a more appropriate or more fitting metaphor for potential disruption than you driving down the street and seeing English teaching skyscrapers and saying to yourself, we're going to go we're going to go disrupt that with people that are teaching that don't speak English, by the way. Right. That is crazy. So, OK, you realize South Korea is your target market. by the way you said you had a i want to back up a little bit you said you had a friend as a tour guide like what was this you did you have like a person in each of these cities like giving you a sales pitch for why their country should be speaks target market like explain that yeah some of them i did some of them in some way or another i just like you know it's like um andrew's parents are from taiwan so like when i was in taiwan his parents were showing us around in korea it's actually a really funny story as well.

53:11This is where this is the single biggest ROI I got. I guess I met my first co-founder as well, but the second or equally positive ROI for going to Harvard for one year, which is that one of my roommates freshman year happened to be someone that was originally from South Korea, lived there until he was like 11 or 12 and then moved to the United States. And so he was like one of these rare individuals, very rare individuals who's like perfectly native in both places and still like has a foot in both. Um, and so I actually, um, he was like a very good friend of mine. Um, and I texted him when I decided I was going to go to Korea and I was like, Hey, how do you feel about going to Korea and visiting your family, all expenses paid trip for, um, you know, on my company.

53:55And he was like, all right, I I'll, I'll take time off right now. This sounds awesome. Um, and, uh, so he flew, flew there with me. Um, And we had an amazing time because we were like really good friends. But he also is this, he was working in New York at the time at Rent the Runway as like a product or data analyst or product analyst. And he was asking, like when we were doing interviews with users, he was like, he would just talk with them for like 10 minutes. And he'd be like, I'd be like, what was that entire conversation? He was like, well, okay, I knew you wanted to ask this, but ultimately, like this is the more interesting thing.

54:29So we just like kind of had this entire conversation. Don't worry, I have notes here. But basically, this is the takeaway. And it was like much higher quality interviews because he was just like so good and so smart. So maybe that was also a reason we went to South Korea, you know, unknowingly, because the interviews were so much higher quality. But yeah, I was in the back of my head. That was like in November. In the back of my head, I was like, OK, that was that was awesome doing that with him. By December, we decided a few weeks later, OK, we're going to go to South Korea. This is it. And so I texted him again and I was like, hey, I have another crazy idea.

55:05What if you quit your job, move to San Francisco in a week and join Speak as basically our first employee and help us figure out South Korea? And he's like, OK, let me think about it. And then I think literally later that day, he was like, all right, I'm in. Let's do this. And he's still at the company to this day. Absolutely like pivotal role. What's his role? Literally, he's done every role under the sun up until now. He figured out our content pedagogy strategy from scratch, became an expert there, did that. Really pushing the boundaries there. I think it's one of our strengths is how much better we are in terms of language acquisition than anyone else.

55:42We've basically invented our own playbook. Now he's doing product work. But he's done so many different things. And he's just classic barrel versus ammunition type. You could just give him something. but like he is like the secret weapon that allowed us to figure out South Korea because the last thing I ever imagined in my life before starting speak was that I would be for like a good three or four years exclusively operating a South Korean consumer app company. Yeah, it's such a crazy concept, right? I mean, especially you're somebody that's built a product in the US before, you know, as you and I've talked about, I'm a former founder of built products here in the US.

56:22And it's so I think it's so important for the founders to be able to see that sort of real-time feedback between what you're shipping and what the users are saying in real time, you can't see that. I mean, I don't think you speak the language, or maybe I missed that part. Maybe you do now. So what is that like? What is it like building a product for another market, especially a startup where that signal is so important? I think especially even consumer, right? Totally. Getting consumers to drop everything in their busy lives and go spend their, not even their money, their time on an experience, especially not like an addictive entertainment experience or a game, but like something that like takes real motivation and willpower.

57:06This is why it took so long to figure out because it was like it was like I like to joke that like every company has like their one unique challenge. They have like the one thing that's like really weird about that company. For us, it was the fact that the founders had never been to South Korea before. We didn't speak the language and we were operating in San Francisco and not in Seoul. That was that was our challenge. And like it very much felt like in many ways it was like one hand tied behind your back when it came to like product iteration and product work and really force you to like do the like do the best practices, not not take the shortcuts, which I think really strengthened us.

57:45And the other really big benefit was that when it came to international expansion, which we have been heavily focused on this year, it's not like most companies when they start international expansion, they're like, oh, OK, we're live in the US. How do we go operate in like the UK or something for us? And it's like a very new muscle to build for us. We've our first market was international expansion. Like we've been an international expansion expert company from the very beginning. So we have four years of doing that before we went to another market. And so, you know, we knew how to hire. We knew the playbook.

58:20We knew like how to localize. We knew how to like design in one language, but then ship something in a different language. We knew how to like do the interviews. Everything like it was something we had like down to a science in terms of like operating and expanding to new markets from the very like from, you know, many years of doing it in South Korea, which is another funny kind of consequence. So once you decide to go all in on South Korea, at this point, you have the product available globally. Now you're just deciding to focus. Is that right? And then and then how long to till things really start working in South Korea?

58:54Yeah, it probably took after being exclusively in South Korea, it probably took like a few months until we started getting like some real pull. And then another like year before we were like, OK, this is something that we definitely want to like the foundation is there and we want to continue to iterate on top of this core like vision and model. And so there were a lot of like little things that all incrementally added up to like a cohesively good experience. But I, yeah, like I think it probably took, you know, at least four more quarters until we were like at that point. And then we started like really thinking about growth and trying to actually scale that up.

59:30Got it. Talk us through the actual speak experience. I mean, right now we've been talking mostly about the insight to be in another market to teach English as a language. But for people that haven't tried the product, and I imagine there will be some that haven't, given that it's, you know, it's targeting teaching English. What is it like compared to, say, Duolingo for learning a non-English language? Yeah. Well, the name speaks, kind of says it all, right? Like the entire idea is we want to build an experience that feels as close to possible. to just kind of having this, you know, across the table experience where you're chatting with this like friendly tutor type and they're just teaching you from the ground up how to become conversationally fluent by getting you to speak out loud a lot, like a ton, like you're speaking more than 50 % of the time.

1:00:22And we're not teaching you grammar. We're not teaching you like weird vocabulary. We're literally just going to focus on teaching you the like core building blocks of actual fluency as if you were in the country or you were like a kid learning for the first time. Like we're just exposing you to the language. And instead of, for example, teaching you, you know, like a classic thing in America or in any language school or, you know, classroom is like the idea of like, OK, we're going to teach you how to conjugate this verb in every single form. And then we're going to teach you the present tense and then the past tense and then the future tense and the like X, Y, Z grammar.

1:00:59And you're going to learn every single one of those forms and memorize it. And then we're going to teach you some vocabulary. And then you're going to be able to say sentences like, I eat at the library, I ate at the library, you know, like sentences that you will never use in real life at all. And for us, we're like, we're not going to teach you any of that. We're going to teach you like the, the three most common ways to use the word eat in actual phrases you're going to use. Like, you know, I, I love to eat thing or I what did you have to, you know, for breakfast yesterday? Oh, I had a blank.

1:01:33We're not even teaching you a we're actually teaching you I had because that's the thing that you're actually going to learn and use the high frequency stuff. And we're not teaching you the underlying logic necessarily of the grammar, but we're going to teach you like the phrases and the patterns that you need. So you can be like, I had a or I had the blank with the blank. And then we teach you a bunch of relevant vocabulary words that fit into those sentences. And then we marry it all together by having you learn a bunch of different phrases and patterns that you repeat so many times that it becomes automatic.

1:02:03Those words are now co-located together. So they all roll off the tongue automatically and you're not translating in your head. You're never translating. And then we marry it all together by having now all of a sudden, you know five or six different patterns and phrases and a little bit of vocabulary that fits with those things. And we put you in a situation where those are actually all the building blocks you need to functionally go do something that you need, that you care about. Like, I want to talk about like going out to eat or I want to be at a restaurant and be able to order food. And all of a sudden you realize, oh, wait, I can do that now.

1:02:35I just unlock this like little part of the universe of fluency of this language, like for this specific use case. I can go do that now at a fluent level. It makes so much sense. You know, anytime I've ever learned a language and I'm sure you've had similar experiences, it's about the immersion, right? It's about talking to somebody. It's not about going through those hypothetical exercises like you mentioned. So is the insight here that enabled this basically that you could take large language model, you know, style application like a chat GBT and you could take, you know, really, really great text to speech and you could effectively enable a human style voice voice conversation.

1:03:14Is that really the insight that led to this product breakthrough? Does that make sense? Yeah, I think it's part of it. Different than like a rules based system like I'm sure Duolingo has, right? Right. Yeah, I think that's part of it. I mean, I think the reality is that we knew that we didn't have like the complete answer on day one. But what we knew was that all these models were going to get better and better and better and surpass humans. And we'd be able to use them in a myriad of different ways, but probably much more important than that. And to our approach, I think this is super relevant for anyone doing anything in the generative AI world.

1:03:50And, you know, a lot of people know this intellectually, but they still, I think, make the mistake is like you need to actually solve real problems. And it's not just about a technology in search of a solution. But like for in the case of Speak, a huge insight for us. And we've seen this like this. We've seen this mistake play out with all the other companies that are now trying to do the same thing is it really matters that we've been able to build an expertise and actually cross functionally fuse the machine learning and the technology. piece and the vision part and building that technology with the content approach that I was just talking about, which was hard, hard learned lessons by doing lots and lots of interviews with users, measuring progress over time and just years of iteration and building that expertise and figuring out how to marry it with the machine learning for today and tomorrow.

1:04:36But then also like building novel product experiences that can, that can combine the content and the pedagogy with the machine learning capabilities of today and tomorrow and doing all that together to actually solve people's problems and not become just a you know, obsessed with what we can do with the latest AI demo and just kind of glue it all together into a product that people want to use once because it's flashy on Twitter, but then they never come back and they don't retain. Yeah, there's so much AI tourism happening with products right now, people trying things just because they're doing new things.

1:05:05But to your point, you need to solve real problems for somebody to keep coming back and for you to have that product market fit. So what does product market fit for Speak look like? You know, you know what it feels like? When did you feel it? Maybe a way to answer that would be, give us a sense of at least what you're comfortable sharing, the scale of this thing in South Korea right now. Yeah. Yeah. I mean, I think for me, like at a high level way, I think about it is like, it's like top line growth and retention. Those are like the things that matter. Like if you're growing naturally, you don't, it's not, not because you're like an absolute paid marketing machine or something like that, but it's just kind of happening without, without you doing anything.

1:05:48and people are sticking around and retaining, those are the signs. Those are like pretty much the only scorecard that matter. In terms of like, yeah, where we are today in Korea, like, you know, we're the number one most popular education app there. I think we've had roughly 6 % of the entire population have tried and used their app. Now they're not all paying subscribers yet, but we think we can get there. But yeah, it's quickly becoming like, if it's not already a household name, um it's it's getting there um and we've helped a lot of people um i think which is the most important part like a lot of people have substantially improved it's really cool it's so cool to see that like talk to users um and like see how much progress they've made uh and uh yeah that's awesome so now like you said this year you've been very very focused on expansion in new markets any specific markets that uh worth mentioning yeah like uh there's i I mean, I think it's a remarkably transferable product, which has been awesome to see.

1:06:53And yeah, like Japan has been a really, really interesting market to us. It's in some ways, you know, a lot of similar learnings, just more for Japan. It's bigger, even more intention in some ways, which has been surprising. And then, yeah, generally, like the we're seeing this as universal. Like we're launching we're live in many markets now. We're live in over 20 countries. We want to like double that by the end of the year. And basically the goal is to be globally available as soon as possible. And I do think like the really cool thing about software and like the entire vision here is it's not like we're scaling a tutor marketplace that is, you know, something that is tricky and like just squishy to actually be able to scale the quality across the new markets and figure out those dynamics like a, like a DoorDash or something, but it's just software.

1:07:42It's just a, it's just a way to get people talking. And like this, everyone learns languages and fundamentally kind of the same way you need to just like practice it and use use an experience like speak. So overall, like what's been really encouraging is just like it does feel like anywhere with this fundamental motivation, people are willing to like go give the product a shot. And then if it works for them, they'll stick around. So if you want to be universally available, then I'm guessing at some point you switch to teaching English speakers other languages. Is that is that part of the plan?

1:08:09Absolutely. Yeah. The United States is, even though people here, it's very US-centric, Anglo-centric, you know, it's just an incredible market. There are lots and lots of people here. Duelingo's built a fabulous business here. And so we think there's a huge opportunity here. And, yeah, there'll be more to announce for that very, very soon, I think. Cool. So do you feel like there are opportunities for speak to help teach people things, you know, other areas of education through AI? Or do you feel like speak will always be focused on the namesake, which is speaking and language? I think, like, obviously there's a lot more to be done in the language space, even just like other parts of it, B2B for children in classrooms for really advanced users, et cetera.

1:08:57Lots and lots of things left to do there. But I think that at the same time, one of the one of the like biggest and most exciting opportunities when it comes to like AI and changing the world and like creating real value for the world is education. Education, I think, traditionally has been like, you know, an underperforming, not as exciting area of like venture capital investment and companies in general, like the companies just aren't that exciting. There's not that many success stories. There's a whole graveyard. hard. Actually, like, ironically, the reason why it's so exciting now is also the reason why there's been nothing exciting so far, right?

1:09:36Which is the reason that like technology hasn't disrupted education yet. People are still learning in the same way that they learned basically 100 years ago, basically since the Industrial Revolution, when we invented the modern classroom. And there's just so much more room for improvement. Like, fundamentally, you go back 2000 years, the way that like Plato learned or, you know, any anyone back in ancient Greece was the best way that the way that the most privileged people learned was like a one on one tutor. And so that hasn't changed since then. But the fact remains that like, virtually everyone don't they don't have access to that.

1:10:11So that's the holy grail for education. It's like, it's like one of the only things in education that's been like scientifically actually proven out the, you know, Bloom's two sigma effect in terms of outcomes for personalized education. And so I think like there's just a tremendous opportunity here. I think we're well positioned for it, but I'm also not arrogant enough to think that like we have the right to win here. There's so many, there's gonna be so many interesting things. I'm really fascinated to see what Khan Academy does. I think that's a really cool opportunity for them as well. But yes, we're certainly going to be thinking about all sorts of things outside of language eventually, or at least that's the ambition.

1:10:50If you weren't building Speak, where else do you feel like there's huge opportunity within AI at the moment? Yeah, I think like there's a few different frameworks you can think about here. But like, I think one of the biggest ones here is like thinking about the stakes of a use case. And there's a certain amount of value that's unlocked when you become an assistant. And that's like that's a total category of use case. Copilot for X, that will be a thing for sure. That's the majority of use cases I feel like right now, especially on the enterprise side. I think that fundamentally the most interesting long-term enduring companies, though, are playing in the spaces where the stakes are low enough or the models will get sufficiently good enough to be above the stakes, if that makes sense, such that you can have something that truly replaces a human altogether.

1:11:39The economic, like, you know, GPT-4 isn't that much cheaper right now than a human. And so like when you can really replace the human altogether, that's where you get the true kind of explosive margins of software. And that is where you like really unlock the like real value. And so, you know, what that means, I think people can think about that for themselves. But I would say like the fundamentally like the most interesting things I think are like where that is an opportunity. Not to say that they're like copilot for X is and also very interesting. but that's like one way I would think about it.

1:12:14It's really easy to see like the upside case for that and the way that it'll give so much time and productivity back to people. I mean, how do you feel about the downside cases of that? Like replacing humans net across the board? What will happen in that scenario? I mean, yeah, like I don't think I'm an expert in that. I would just caveat it with that, like really take it with a grain of salt. But if I were to speculate, like I'm fundamentally an optimistic person, I think. I believe that technology historically has almost always been mostly used for good. There's obviously always the flip side of that.

1:12:52I do think there's a danger if we like basically get to AGI too quickly and we don't we're not able to properly handle that. That's that's like the risk that I'm more concerned about. Candidly, I think in terms of like the economics of it, like everything is accelerating. this will happen faster than the last technology wave. But I do believe that fundamentally, we're just in a better world with this technology. We're able to continue to remove the yoke from humans in terms of doing the work that isn't fulfilling for the people and is just purely a paycheck and fuel economic growth and be able to just unlock better jobs.

1:13:30So I fundamentally think that's the case. But I also do know there's totally a flip side here for the people that are you know, maybe later in their careers and less willing to like make a transition. And that's really tricky. I don't have the answers there. But at the same time, this is inevitable. Like, I do think like fighting, fighting against it is is kind of like trying to, you know, fight quicksand or swim up a river that's just flowing the opposite way. Like, it's just it's it's hard to do. And so I think it's it's one of those things where we need to just figure out how to embrace it.

1:14:04You mentioned earlier in the conversation that when you were thinking about Speak, it was pretty clear when it was going to get good enough by feeding it enough data. And you also mentioned early on that one of the coolest things one could do with AI is build effectively sentient AI. I'm sure this question gets asked to you often, but you've been ahead of the curve for so long like when is that going to happen when is sentient like when is like basically agi gonna happen yeah exactly yeah um well the print i feel like you know the the classic thing here is like it's gonna be like five to ten years away or or beyond um and um i i like i think it's very tricky i think it's like it's very unclear what that even means can GPT-4 pass the Turing test?

1:14:57Yeah, absolutely. That's what we thought. That's what we thought AGI was up until recently. Are we going to continue to move the goalposts? Probably. Human level intelligence is a little bit arbitrary. It's important in the narrow case. I'm not sure it's important in the macro case. I think the bigger concern is if there's a runoff where it just continues to accelerate really quickly and exponentially. And that's the super intelligence scenario. I think it's pretty clear that that's probably not that likely. At least it's not going to be like, oh, you were going to blink our eyes and it's going to happen.

1:15:29I think we're fundamentally very clearly constrained by NVIDIA, basically, and GPUs. And five years ago, when everyone was thinking about reinforcement learning, I was more concerned about that. But if it's going to happen with this current... And also, by the way, I think there's some hard limitations. The transformer, we're squeezing a lot of juice out right now, but we're going to need to make some breakthroughs. to get to the next level. I'm pretty convinced of that. So that's why it's indeterminate how long it will take because there will need to be some fundamental new breakthroughs, I believe.

1:16:00But yeah, I think fundamentally, like five years ago, when we, you know, when everyone was focused on reinforcement learning, it was less clear what that like, what the kind of like limiting factors would be for accelerations. And I think now it's like pretty clear compute and energy are those. So like, it's only going to get so much, like NVIDIA will only get so much better at making really, really good A100s and stuff over the next time. And I don't think we're close yet in terms of having the compute to be able to easily do, like to be able to get to the point where that's the case. When compute doesn't become the limiting factor, that's when I would be more nervous about this happening and having the runaway scenario.

1:16:41But as long as that's the case, look, you can monitor how much electricity a certain, you know, super, uh, supercomputer cluster is using. Um, and like, that's a pretty good proxy for kind of like the level of, of complexity and reasoning that the system can use. So, um, so that's why I think like, I do think like opening eyes approach here is very reasonable. And I, I do think they genuinely care about AI safety and all of that. And I, I think like, I'm pretty sure they view it not that differently than that as well. Connor, anyone listening to this interview, I'm sure, finds Speak to be fascinating.

1:17:17So I have to ask, is Speak hiring? Absolutely. Thank you for asking that. Yeah, I mean, basically, if there's anyone that's super interested in this, especially if you're technical, we'd love to talk to you. Definitely anyone on the kind of engineering product design, machine learning stack, we're looking to hire across the board there. And like, you know, definitely the right kind of ops hire or growth hire as well. So yeah, definitely, please reach out. You can just go to speak.com slash careers to find out more and get in touch with us. And for people who are in the markets where you're currently focused, how can they try the product?

1:17:58Yeah, just go to our website, speak.com and you can download the app from there. And always love feedback, especially if it's in our new markets. I'm sure there's things we can do to improve. How can they give the feedback? There's an easy way to give feedback directly in the app, or you can just shoot us a line at contact at speak.com. Awesome. Or feedback at speak.com. Yeah. Awesome. Connor, thank you so much. This was an incredible conversation. Appreciate you giving us all this time and hope to have you on again sometime. Yeah. Thanks for having me. This was awesome. Great questions. Thank you so much for listening to my conversation with Connor.

1:18:36Follow Generative Now wherever you get your podcast on Apple Podcasts and Spotify and YouTube and wherever else you listen. Also, if you enjoyed the episode, please do us a favor and rate and review it. It really, really helps. If you wanna hear more from me or Lightspeed, follow me at McNano on X or Twitter or LinkedIn, Instagram, everywhere. And if you wanna follow us at Lightspeed, you can find us at Lightspeed VP on all those same platforms. Thanks again to Connor for the awesome interview. and Generative Now is produced by Lightspeed in partnership with Pod People. Special thanks to everyone on our production team for making this podcast possible.

1:19:14We will be back next week with another awesome conversation, so don't miss it. See you next time.